We are in an age in which more and more application areas for machine learning are found. The possibilities are also being explored in the field of finance. However, unlike a simple method like linear regression, deep learning methods are rather complex and not tangible for potential investors. This thesis deals with the application of neural networks, their explainability, and the trading of Bitcoin. The theory of the neural networks used and the novel explainability aspect is explained, furthermore, there is a brief insight into the Bitcoin theory. We use the autocorrelation of log returns to find the input layer of the optimal network architecture. The number of hidden layers and neurons is then determined by quantitatively comparing the error function MSE and the Sharpe ratio for each possible combination. An XAI application, Linear Parameter Data (LPD), is used to determine in which phases the neural network is reliable. Combined with a volatility prediction from a GARCH model, a trading strategy is implemented and improved with the addition of the cryptocurrency Ether. The results are promising, as the inclusion of LPD lead to an added value in our scope. Furthermore, the addition of trades based on Ether enables to even outperform a Bitcoin buy-and-hold strategy.
Authors
- Geeler Ken
- Bühler Pascal
- Rieser Philipp
Supervisor
- Wildi Marc